Object Detection Using Frame Feature Map Pooling for Slow-Moving Targets
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current object detection methods, particularly those using dynamic vision sensors, struggle to accurately detect slow-moving objects due to insufficient information in single frame images captured by these sensors, leading to inaccurate detection and processing inefficiencies.
Innovation Solution
The proposed method involves acquiring a sequence of frames, extracting feature maps, pooling them with information from previous frames using a weight image calculated through a multilayer perceptron neural network, and recursively applying this process to generate a pooled feature map that combines information from multiple frames for more accurate object detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If single frame image detection is used with DVS camera, then fast moving objects can be detected, but slow moving objects cannot be detected accurately
Solution Approach 1:
The patent merges feature maps from multiple consecutive frames by pooling operations to create a combined feature representation. This allows slow-moving objects that appear subtly across frames to be detected through accumulated evidence, while fast-moving objects are still detectable in individual frames within the sequence.
Solution Approach 2:
The system performs preliminary feature extraction on each frame individually before pooling them together. This preliminary action ensures that even fast-moving objects that may only appear in one frame have their features extracted and are preserved in the final pooled feature map for detection.
2Measurement precision
If multiple frames are pooled for detection, then detection accuracy improves, but processing complexity increases
Solution Approach 1:
The patent segments the processing into distinct stages: feature extraction for each frame, pooling operation to combine features, and final detection. This segmentation allows efficient processing by handling each frame's features independently before combining them, reducing overall computational complexity compared to processing all frames simultaneously.
Solution Approach 2:
The system extracts features for all frames in the sequence (excessive action) but only pools and processes the essential combined features for detection (partial action). This avoids the complexity of processing every detail of every frame while still achieving improved detection accuracy through multiple frames.
3Loss of information
If feature maps from multiple frames are combined, then information completeness increases, but computational load increases
Solution Approach 1:
The patent extracts only the essential features from each frame using feature extraction operations before pooling. This extraction process removes redundant information while preserving critical object characteristics, allowing multiple frames to be combined without proportionally increasing computational load.
Solution Approach 2:
The system performs feature extraction on all frames (excessive action to ensure no information is missed) but applies pooling operations only to the extracted feature representations rather than raw image data (partial action to reduce computational load).
Data Source
AI summary
An object detection method and a neural network system for object detection are disclosed. The object detection method acquires a current frame of a sequence of frames representing an image sequence, and extracts a feature map of the current frame. The extracted feature map is pooled with information of a pooled feature map of a previous frame to thereby obtain a pooled feature map of the current frame. An object is detected from the pooled feature map of the current frame. A dynamic vision sensor (DVS) may be utilized to provide the sequence of frames. Improved object detection accuracy may be realized, particularly when object movement speed is slow.